Learning from Small Datasets with Hybrid Swin-Transformer in Industrial CT: A Case Study on Boxed Shoes

Transformer-based architectures excel at capturing global context, making them promising candidates for segmenting complex industrial Computed Tomography data. However, they typically require large datasets and substantial memory resources. In this work, we adapt a hybrid Swin Transformer–CNN model to small, volumetric industrial CT datasets. We achieve this by reducing the number of parameters and adding an additional convolutional layer and a skip connection before the transformer. Thereby, the model segments patches up to 2563 voxels on 24 GB GPU memory. Trained on 25 volumetric scans of boxed shoes, the proposed model achieves F1 -scores of 85.2% for shoe segmentation and 73.4% for the inner volume. For component-level segmentation, the scores are 39.6% for the insole, 71.7% for the outsole, and 62.0% for the shoe upper. Compared to a Residual SE U-Net baseline, the hybrid model performs comparably in global segmentation tasks but shows a larger drop for fine-grained, local structures. This work demonstrates that even lightweight hybrid transformers can provide robust generalization when global context is required, making them suitable for small industrial CT datasets.